Fast Machine Learning for Science Conference 2026

Maskformer models for particle reconstruction in high-granularity calorimeters

Not scheduled

12m

Presentation Contributed Talks

Speaker

Matteo Marchegiani (Carnegie-Mellon University (US))

Description

The calorimeters employed in large experiments at the Phase-1 LHC were designed to cope with the radiation levels and hit multiplicities expected in proton-proton collisions with an average of 20–50 interactions per bunch crossing. However, the High-Luminosity LHC (HL-LHC) will operate with pileup levels of up to 200 interactions per crossing. The radiation environment is especially challenging in the detector endcaps, where limited longitudinal and transverse segmentation complicates particle reconstruction in the forward region.

These effects can be mitigated through the use of highly granular calorimeters that combine precise spatial and timing information, such as the CMS High-Granularity Calorimeter (HGCAL). However, their increased complexity requires the development of new algorithms for fast and efficient particle reconstruction.

Machine learning (ML) provides a natural approach to particle reconstruction in dense environments, leveraging raw detector-hit information to reconstruct particles in parallel while intrinsically accounting for overlapping showers. Transformers have revolutionized natural language processing and computer vision, and their application to experimental particle physics has grown rapidly in recent years due to their strong performance. In particular, MaskFormers have demonstrated outstanding results in image object detection tasks. Unlike graph-based clustering algorithms relying on Graph Neural Networks (GNNs), transformers learn clustering through self-attention over the full detector-hit representation, enabling the identification of long-range correlations without requiring a predefined graph structure. Furthermore, particle properties such as energy and position can be regressed directly from the latent representation. From a computational perspective, MaskFormers mitigate the quadratic complexity of transformers through masked attention, restricting attention calculations to geometrically relevant subsets of detector hits while maintaining performance.

A particle reconstruction algorithm for high-granularity calorimeters based on the MaskFormer architecture is presented, and its performance in terms of energy and position resolution is evaluated using simulated events.